Farzad Habibi: Metastable Failures in Consensus Algorithms

Abstract

Metastable failure is a recent abstraction of a pattern of failures in distributed systems. A metastable failure is characterized as “permanent overload with an ultra-low goodput.”
Prior research has proposed a framework for understanding metastable failure and has observed various cases of such failures in real-world settings.
In this talk, we discuss the challenge of metastable fault tolerance in replication systems, focusing specifically on a basic problem in distributed systems—consensus. Consensus is a basic building block for many distributed systems and protocols which means that such a focus would have an impact on a large class of systems.
The main points covered in this talk include (1) an introduction to metastable failures, (2) a comprehensive analysis of various metastable failure cases in replication systems, (3) a reproduction of a metastable failure case study, (4) modeling of these failures using queuing theory, and (5) a machine learning approach for predicting metastable failures.

Bio:

Farzad Habibi is a second-year Ph.D. student in Computer Science at the University of California, Irvine. He earned his Bachelor’s degree in Computer Engineering from the University of Tehran. Prior to his current research focus, Farzad explored the resiliency of blockchain systems in the face of network partitioning. His current research endeavors are centered around investigating and addressing metastable failures in distributed systems.